The Brain’s Concepts

The Role of the Sensory-Motor System in Reason and Language

George LakoffUniversity of California, Berkeley

(with Vittorio Gallese)

With Thanks to

The Neural Theory of Language Group

International Computer Science Institute

University of California, Berkeley

Especially Jerry Feldman, Srini Narayanan,Lokendra Shastri, and Nancy Chang.

http://www.icsi.berkeley.edu/NTL

What Concepts Are: Basic Constraints

Concepts are the elements of reason, and

constitute the meanings of words and linguistic expressions.

The Traditional Theory

Reason and language are what distinguish human beings from other animals.

Concepts therefore use only human-specific brain mechanisms.

Reason is separate from perception and action, and does not make direct use of the sensory-motor system.

Concepts must be “disembodied” in this sense.

We Claim

Human concepts are embodied. Many concepts make direct use of the sensory-motor capacities of our body-brain system.

Many of these capacities are also present in non-human primates.

One example, the concept of grasping, will be discussed in detail.

Amodality

The traditional theory implicitly claims that even action concepts, like grasp, do not make use of the sensory-motor system. As a concept, even grasp must be disembodied.

Thus, it is claimed that the concept grasp is amodal. Since it is a concept, it must be modality-free, even if it designates an action in a specific modality.

Concepts Are:

•Universal: they characterize all particular instances; e.g., the concept of grasping is the same no matter who the agent is or what the patient is or how it is done.

•Stable.

•Internally structured.

•Compositional.

•Inferential. They interact to give rise to inferences.

•Relational. They may be related by hyponymy,

antonymy, etc.

•Meaningful.

•Independent of the words used to express them.

Concepts may be either

‘concrete’ (sensory-motor)

or

‘abstract’ (not sensory-motor).

Basic Ideas

•Multimodality — Permits universality

•Functional Clusters — High-level, function as

conceptual units

•Simulation — Necessary for meaningfulness

and contextual inference

•Parameters — Govern simulation, strict

inference, link to language

Multimodality

The action of grasping is not amodal, but multi-modal in a way that makes for universality.

Functional Clusters

Functional clusters form high-level units — with the internal relational structure required by concepts.

There are two types: Local clusters and Network clusters.

Multi-modality is realized in the brain through network clusters, that is, parallel parietal-premotor networks.

Network clusters are formed by interconnected local clusters of neurons, like canonical and mirror neurons.

Simulation

To understand the meaning of the concept grasp, one must at least be able to imagine oneself or someone else grasping an object.

Imagination is mental simulation, carried out by the same functional clusters used in acting and perceiving.

The conceptualization of grasping via simulation therefore requires the use of the same functional clusters used in the action and perception of grasping.

Simulation and Enactment

Visual imagination uses part of the same neural substrateas vision.

Motor imagination uses part of the same neural substrate is motor action.

Since you can understand a concrete concept like grasping only if you can imagine doing it or observing it,

the capacity for mental simulation is taken as the basis formeaningfulness.

Thus, action and observation provide the basis for meaningfulness in NTL.

Parameters

All actions, perceptions, and simulations make use of parameters and their values. Such neural parameterization is pervasive.

E.g., the action of reaching for an object makes use of the parameter of direction; the action of grasping an object makes use of the parameter of force.

The same parameter values that characterize the internal structure of actions and simulations of actions also characterize the internal structure of action concepts.

Structured Neural Computation in NTL

The theory we are outlining uses the computational modeling mechanisms of the Neural Theory of Language (NTL).

NTL makes use of structured connectionism (Not PDP connectionism!).

NTL is ‘localist,’ with functional clusters as units.

Localism allows NTL to characterize precise computations, as needed in actions and in inferences.

Because it uses functional clusters, NTL is not subject to the “grandmother cell” objection.

Advantages of Structured Connectionism

Structured connectionism operates on structures of the

sort found in real brains.

From the structured connectionism perspective, the

inferential structure of concepts is a consequence of the

network structure of the brain and its organization in

terms of functional clusters.

Structured Connectionism comes with:

•A dynamic simulation mechanism that adapts

parameter values to situations.

•A neural binding theory.

•A spreading-activation probabilistic inference

mechanism that applies to functional clusters.

These jointly allow for the modeling of both sensory-

motor simulations and inference.

In NTL, there are fixed structures called schemas.

For example, a schema that structures an action hasan internal structure consisting of Roles, Parameters,and Phases.

The ideas of Multimodality, Functional Clusters, Simulation, and Parameters allow us to link NTL, with structured connectionism, to neuroscience.

The Neuroscience Evidence Shows

In the sensory-motor system, it is possible to

characterize these aspects of concepts:

•Universality

•Semantic Role Structure

•Aspectual Structure (Phases)

•Parameter Structure

The Concept

Of

Grasping

Universality Is Achieved by MultiModality

Multimodal functional clusters for an action like grasping fire when:

•Grasping is performed, observed, imagined, inferred, or heard;

•The grasping is of any type, done by any agent, on any object,in any manner, and in any location.

In showing such multimodality for a functional cluster, we are showing that the functional cluster plays the conceptual role of universality.

Multi-Modal Integration

The premotor cortex is not a uniform field, but a mosaic

of functionally distinct areas (F1 to F7).

Each of these premotor areas is reciprocally connected

with distinct regions of the posterior parietal cortex.

The premotor cortex is part of a series of parallel

functional network clusters.

Multi-Modal Integration

Cortical premotor areas are endowed with sensory properties.

They contain neurons that respond to visual, somatosensory, and auditory stimuli.

Posterior parietal areas, traditionally considered to process and associate purely sensory information, also play a major role in motor control.

Rizzolatti et al. 1998

A New PictureA New Picture

The fronto-parietal networks

Rizzolatti et al. 1998

Area F5Area F5

Three classes of neuronsThree classes of neurons:

-Motor General Purpose neurons-Motor General Purpose neurons

-Visuo-Motor neurons:-Visuo-Motor neurons:

-Canonical neurons-Canonical neurons

-Mirror neurons-Mirror neurons

Area F5

General Purpose Neurons:General Grasping

General Holding

General Manipulating

AA Grasping with the mouth

BB Grasping with the cl. hand

CC Grasping with the ipsil. hand

General Purpose Neurons in Area F5General Purpose Neurons in Area F5

(Rizzolatti et al. 1988)

General Purpose Neurons Achieve

Partial Universality: Their firing correlates with a goal-oriented action of a general type, regardless ofeffector or manner.

F5c-PFF5c-PF

Rizzolatti et al. 1998

The F5c-PF circuit

Links premotor area F5c and parietal area PF (or 7b).

Contains mirror neurons.

Mirror neurons discharge when:

Subject (a monkey) performs various types of goal-related hand actions

and when:

Subject observes another individual performing similar kinds of actions

Area F5cArea F5c

Convexity region of F5:

Mirror neurons

F5 Mirror NeuronsF5 Mirror Neurons

Gallese and Goldman, TICS 1998

Strictly congruent mirror neurons (~30%)

(Rizzolatti et al. Cog Brain Res 1996)

Category Loosening in Mirror Neurons (~60%)

(Gallese et al. Brain 1996)

PF Mirror NeuronsPF Mirror Neurons

(Gallese et al. 2002)

Umiltà et al. Neuron 2001

A (Full vision)A (Full vision)

B (Hidden)B (Hidden)

C (Mimicking)C (Mimicking)

D (HiddenMimicking)D (HiddenMimicking)

Like humans, monkeys can also infer the goal of an action, even when the visual information about it is incomplete.

F5 Audio-Visual Mirror NeuronsF5 Audio-Visual Mirror Neurons

Kohler et al. Science (2002)

Somatotopy of Action ObservationSomatotopy of Action Observation

Foot ActionFoot Action

Hand ActionHand Action

Mouth ActionMouth Action

Buccino et al. Eur J Neurosci 2001

The Mirror System in HumansThe Mirror System in Humans

BA6

The Simulation Hypothesis

How do mirror neurons work?

By simulation.

When the subject observes another individual doing an action, the subject is simulating the same action.

Since action and simulation use some of the same neural substrate, that would explain why the same neurons are firing during action-observation as during action-execution.

Mirror Neurons Achieve

Partial Universality, since they code an action regardless of agent, patient,modality (action/observation/hearing),manner, location.

Partial Role Structure, since they codean agent role and a purpose role.

The Agent Role: In acting, the Subject is an agent of that action.In observing, the Subject identifies the agent ofthe action as having the same role as he haswhen he is acting – namely, the agent role.

The Purpose Role: Mirror neurons fire only forpurposeful actions.

Mirror Neurons Achieve

Category tightening and loosening

Limited Prototype Structure

F5ab-AIP

The F5ab-AIP circuit

Links premotor area F5ab and parietal area AIP.

Transforms intrinsic physical features of objects (e.g., shape, size)

into hand motor programs required to act on them

Examples:

Manipulate objects, grasp them, hold them, tear them apart.

Area F5abArea F5ab

Bank region of F5:

Canonical neuronsCanonical neurons

Murata et al. J Neurophysiol. 78: 2226-2230, 1997

F5 Canonical NeuronsF5 Canonical Neurons

F5 Canonical NeuronsF5 Canonical Neurons

Murata et al. J Neurophysiol. 78: 2226-2230, 1997

The Simulation Hypothesis

How Do Canonical Neurons Work?

By Simulation.

The sight of a graspable object triggers the simulation of grasping.

Since action and simulation use some of the same neural substrate, that would explain why the same neurons are firing during object-observation as during action-execution.

Canonical Neurons Achieve

Partial Universality, since they code an action regardless of patient,manner, and location.

Partial Role Structure, since they codea patient role and a purpose role.

The Patient Role: Canonical neurons firein the presence of an appropriate patientfor a given action.

The Purpose Role: Canonical neurons fireonly for purposeful actions.

F4-VIPThe F4-VIP Network Custer

The F4-VIP Circuit

Links premotor area F4 and parietal area VIP.

Transforms the spatial position of objects in peri-personal space

into motor programs for interacting with those objects.

Examples:

Reaching for the objects, or moving away from them

with various parts of your body such as the arm or head.

Area F4Area F4

Arm reaching

Head turning

Somato-Centered Bimodal RFs in area F4Somato-Centered Bimodal RFs in area F4

(Fogassi et al. 1996)

(Graziano et al. 1999)

Somato-Centered Bimodal RFs in area VIPSomato-Centered Bimodal RFs in area VIP

(Colby and Goldberg 1999)

Somato-Centered Bimodal RFs in area F4

(Fogassi et al. J Neurophysiol 1996)

The Simulation Hypothesis

How Do Action-Location Neurons Work?

By Simulation.

The sight or sound of a possible target location inperi-personal space triggers the simulation of appropriate actions toward that location.

Since action and simulation use some of the same neural substrate, that would explain why the same neurons are firing during location-perception as during action-execution.

Action-Location Neurons Achieve

Partial Universality, since they code an action regardless of patient.

Partial Role Structure, since they codeLocation.

Evidence in Humans for Mirror, Canonical, and Action-Location

Neurons

Mirror: Fadiga et al. 1995; Grafton et al. 1996;Rizzolatti et al. 1996; Cochin et al. 1998;

Decety et al. 1997; Decety and Grèzes 1999;Hari et al. 1999; Iacoboni et al. 1999;

Buccino et al. 2001.

Canonical: Perani et al. 1995; Martin et al.1996; Grafton et al. 1996; Chao and Martin 2000.

Action-Location: Bremmer, et al., 2001.

MULTI-MODAL INTEGRATION

The premotor and parietal areas, rather than havingseparate and independent functions, are neurally integratednot only to control action, but also to serve the function ofconstructing an integrated representation of:

(a) Actions, together with (b) objects acted on, and (c) locations toward which actions are directed.

In these circuits sensory inputs are transformed in order toaccomplish not only motor but also cognitive tasks, such asspace perception and action understanding.

Phases

Area F5 contains clusters of neurons that control distinctphases of grasping: opening fingers, closing fingers.

Jeannerod, et al., 1995; Rizzolatti, et al., 2001.

Summary

Jointly, these functional clusters in the sensory-motor systemcharacterize the following conceptual properties of grasping:

•Stability

•Universality: Covers all particulars

•Internal Structure:

Semantic Roles

Phases (Aspectual Structure)

•Meaningfulness

•Independence of linguistic expression

Summary

In NTL, structured connectionist mechanisms apply tounits modeling functional clusters.

Compositionality is modeled via neural binding.

Inference is modeled via structured connectionistmechanisms for: binding, spreading activation inference, andsimulation.

Conclusion 1The Sensory-Motor System Is Sufficient

For at least one concept, grasp, functional clusters, as

characterized in the sensory-motor system and as modeled

using structured connectionist binding and inference

mechanisms, have all the necessary conceptual properties.

Conclusion 2The Neural Version of Ockham’s Razor

Under the traditional theory, action concepts have to be disembodied, that is, to be characterized neurally entirely outside the sensory motor system.

If true, that would duplicate all the apparatus for characterizing conceptual properties that we have discussed. Unnecessary duplication of this sort is highly unlikely in a brain that works by neural optimization.

How does NTL fit the Neuroscience?

Actions in NTL

For each type of action there is a Fixed Schema, consistingof types of fixed parameters; for example:

•Role Parameters, like Agent and Patient •Phase Parameters, like Initial and Final State•Manner Parameters, like Degree of Force and Direction

Grasp Schema

Roles: Action, Agent, Patient, LocationManners: Force, Type of Grip; Effector UsedPhases:

Initial State:: Object Location: Within Peri-personal Space

Starting Transition:: Reaching, with Direction: Toward Object

Location; Opening Effector

Central Transition:: Closing Effector, with Force: A function

of Fragility and Mass of Patient

Goal Condition:: Effector Encloses Object, with Manner: (a

grip determined by parameter values and situational

conditions)

Final State:: Agent In-Control-of Object

Fitting The Grasp Schema to the Neuroscience of Grasping

A Fixed Schema Is a Network of Functional ClustersEach Parameter Is a Functional Cluster of neuronsEach Parameter value Is either

A firing pattern over a functional cluster, orA neural binding to another functional cluster,

as when the role Agent is bound to a particularactor in context.

An Executing Schema (X-schema) Is a neural circuit connecting the parameters of the fixed schema so that they can dynamically coordinate firing over time andadapt their values over time to inputs from context.

Note!

The same neurons that define the fixed schema are the neurons subject to dynamic, contextually adjusted activation by the executing schema during performance, observation, and imagination.

Schemas are not like logical conditions. They run bodies — as well as they can, in real time adjusting to real conditions.

Other Differences From Traditional Accounts of Concepts

•Not Necessary and Sufficient Conditions

•Not Representational

•Not Symbolic

Not Necessary and Sufficient Conditions

•The activation of functional clusters is not all-or none; there

are degrees.

•There are variations on schemas, as when certain phases

are optionally left out.

•There are extensions of schemas; for example, extensions

from the prototoype and metaphorical extensions.

Not Representational

We conceptualize the world on the basis of the way we experience it; e.g., color is not in the world, nor is heat.

Since our experience is a function of our bodies, brains, and our physical and social environment, so are our concepts.

Since our experience comes through our physical nature — our bodies, brains, and physical functioning — so our concepts are physical in nature.

They are physical brain structures that, when activated, result in creative understandings shaped by the peculiar character of our bodies, brains, and lived experiences.

Not Symbolic

.Note that we have written down symbols (e.g., Final State) as our notation for functional clusters.

This does NOT mean that we take functional clusters themselves to be symbolic. We only use symbols because we have to write things down.

The symbols are only our names for functional clusters, which, as we have seen, are made of neurons, though they function — from a computational modeling point of view — as units.

Language is Multi-Modal, Not Modular

Concepts form the most interesting part of language, the meaningful part.

Many concepts, which are part of language, are inherently multi-modal, exploiting the pre-existing multi-modal character of the sensory-motor system.

It follows that there is no single “module” for language — and that human language makes use of mechanisms present in nonhuman primates.

What About Abstract Concepts?

Abstract Concepts

Not all concepts are about physical things or what we do with our bodies.

Some are about emotions, like love.

Others are even less concrete, like freedom.

Conceptual Metaphor ProvidesEmbodied Reasoning For Abstract

Concepts

Virtually all abstract concepts (if not all) have conventional metaphorical conceptualizations — normal everyday ways of using concrete concepts to reason systematically about abstract concepts.

Most abstract reasoning makes use of embodied reasoning via metaphorical mappings from concrete to abstract domains

What Are Conceptual Metaphors?

In NTL, conceptual metaphors are structured connectionist “maps” — circuits linking concrete sourcedomains to abstract target domains.

In the fit of NTL to Neuroscience, such metaphorical maps would be neural circuits in the brainlinking sensory-motor regions to other regions.

We claim therefore that, in such cases, the sensory-motor system is directly engaged in abstract reasoning.

Metaphorical Grasping

There is a conceptual metaphor, Understanding Is Grasping, according to which one can grasp ideas.

Reasoning patterns about physical grasping can be mapped by conceptual metaphor onto abstract reasoning patterns.

One can begin to grasp an idea, but not quite get a hold of it.

If you fail to grasp an idea, it can go right by you — or over your head!

If you grasp it, you can turn it over in your mind.

You can’t hold onto an idea before having grasped it.

The Sensory-Motor System in Abstract Reasoning

We have argued that the physical Grasping Schema is realized in the sensory-motor system, and that its inferences are carried out imaginatively in sensory-motor simulation.

At least some of these inference patterns are used metaphorically to do abstract reasoning about understanding.

If our analysis is correct, then the sensory-motor system is directly engaged in abstract reasoning.

Cogs

The exploitation of “general” sensory-motor mechanisms

for abstract reasoning

and grammar

Premotor Versus Motor Cortex

Whenever we perform a complex motor movement, such as picking up a glass and taking a drink, at least two distinct parts of the brain are activated:

The motor cortex, where there are neural ensembles that control “motor synergies” — relatively simple actions like opening or closing the hand, flexing or extending the elbow, turning the wrist, and so on.

Complex motor schemas, however, are carried out by neural circuitry in the pre-motor cortex, circuitry connected via neural bindings to the appropriate synergies in the motor cortex.

In picking up a glass and taking a drink, both pre-motor cortex and motor cortex are activated, as are binding connections between them.

The Controller X-Schema

In modeling complex premotor action schemas, Narayanan made a remarkable discovery.

All complex premotor schemas are compositions of a single type of structure.

He then showed that the same neural computational structure, when disengaged from specific motor actions, can characterize aspect (that is, event structure) in the world’s languages. When dynamically active, this structure can compute the logic of aspect.

Narayanan called this structure the “Controller X-schema.”

The Structure of the Controller X-Schema

•Initial State•Starting Phase Transition•Precentral State•Central Phase Transition (either instantaneous,

prolonged, or ongoing)•Postcentral State*•Ending Phase Transition•Final State

Postcentral Options: *A check to see if a goal state has been achieved *An option to iterate or continue the main process *An option to stop/resume

-Narayanan, 1997

The Controller X-Schema as a Computational Model

The Controller X-Schema is implemented computationally using Petri Nets that have been greatly revised and extended to closely approximate neural systems.

Narayanan has developed his program to be a general mechanism for imaginative simulation.

The computational model is intended to be mapped onto neural structures so that we can speak of neural Controller X-Schemas with the following properties.

The Properties ofA Neural Controller X-Schema

•It is a neural structure that is “general” in the sense that it can be bound via connections to different specific sensory-motor structures elsewhere in the brain.

•When those connections are deactivated, it can be connected to other regions of the brain and perform abstract reasoning.

•In its reasoning mode, it characterizes the semantics of a portion of grammar (e.g., aspect and its logic).

•The inference patterns it characterizes are “general,” in that they can apply to a wide range of specific concepts.

I will call any neural structure with such properties a “Cog.”

Other Cogs

Other examples of Cogs include primitive image-schemas —

e.g., Containers, Source-Path-Goal, Contact, Rotation, Front-

Back, Up-Down— as well as Talmy’s force dynamic schemas,

enumeration schemas (used in subitizing), and so on.

All of these can be bound to a wide range of specific sensory-

motor details, can be used in reasoning, and can characterize

the meanings of grammatical constructions.

Other Uses of Cogs

Linking metaphors that join different mathematical domains are Cog-to-Cog mappings.

e.g., Numbers Are Points on a Line

Cogs characterize form in art.

Dissociative learning is the inhibition of connections between Cogs and specific details.

The Sensory-Motor Nature of Cogs

The primary function of Cogs is a sensory-motor function.

Both evolutionarily and developmentally, Cogs first functionto structure our embodied sensory-motor interactions in the

world.

That function is not lost. Cogs continue in their sensory-motorfunction.

The sensory-motor characteristics of Cogs are exploited in reason, language, mathematics, and art — the highest of

human cognitive functions.

All of these make direct use of the sensory-motor neuralsubstrate!